Evaluation of Dimensional Reduction Methods on Urban Vegegation Classification Performance Using Hyperspectral Data

Abstract

In the context of urban vegetation, hyperspectral imagery allows to discriminate biochemical properties of land surfaces. In this study, we test several dimension reductions to evaluate capacities of hyperspectral sensors to characterize tree families. The goal is to evaluate if a selection of differentiated and uncorrelated vegetation indices is an efficient method to reduce the dimension of hyperspectral images. This method is compared with conventional MNF and ACP approaches, and assessed on tree vegetation classifications performed using SVM classifier on two datasets at 4m and 8m spatial resolution. Results show that MNF combined with SVM classification is the better method to reduce hyperspectral dimension.